floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request #3493 from martin-frbg/casts+cleanup 

WIP casts and cleanups
This commit is contained in:
Martin Kroeker 2022-02-06 23:55:06 +01:00 committed by GitHub
parent aec32e5bd4 7b146e590c
commit 7656aba00e
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
19 changed files with 55 additions and 38 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
cpuid_x86.c
View File

@ -323,9 +323,11 @@ int get_vendor(void){
int get_cputype(int gettype){ int get_cputype(int gettype){
int eax, ebx, ecx, edx; int eax, ebx, ecx, edx;
/*
int extend_family, family; int extend_family, family;
int extend_model, model; int extend_model, model;
int type, stepping; int type, stepping;
*/
int feature = 0; int feature = 0;
cpuid(1, &eax, &ebx, &ecx, &edx); cpuid(1, &eax, &ebx, &ecx, &edx);
@ -428,7 +430,8 @@ int get_cacheinfo(int type, cache_info_t *cacheinfo){
cpuid(0, &cpuid_level, &ebx, &ecx, &edx); cpuid(0, &cpuid_level, &ebx, &ecx, &edx);
if (cpuid_level > 1) { if (cpuid_level > 1) {
int numcalls =0 ; int numcalls;
cpuid(2, &eax, &ebx, &ecx, &edx); cpuid(2, &eax, &ebx, &ecx, &edx);
numcalls = BITMASK(eax, 0, 0xff); //FIXME some systems may require repeated calls to read all entries numcalls = BITMASK(eax, 0, 0xff); //FIXME some systems may require repeated calls to read all entries
info[ 0] = BITMASK(eax, 8, 0xff); info[ 0] = BITMASK(eax, 8, 0xff);
@ -1637,7 +1640,6 @@ int get_cpuname(void){
else else
return CPUTYPE_BARCELONA; return CPUTYPE_BARCELONA;
} }
break;
case 10: // Zen3 case 10: // Zen3
if(support_avx()) if(support_avx())
#ifndef NO_AVX2 #ifndef NO_AVX2
@ -2193,7 +2195,6 @@ int get_coretype(void){
else else
return CORE_NEHALEM; return CORE_NEHALEM;
#endif #endif
break;
case 7: case 7:
if (model == 10) if (model == 10)
@ -2582,4 +2583,4 @@ void get_sse(void){
if (features & HAVE_FMA3 ) printf("HAVE_FMA3=1\n"); if (features & HAVE_FMA3 ) printf("HAVE_FMA3=1\n");
} }
//} //}

40
driver/others/blas_server.c
View File

@ -209,7 +209,8 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
/* REAL / Double */ /* REAL / Double */
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, double, void (*afunc)(BLASLONG, BLASLONG, BLASLONG, double,
double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG,
double *, BLASLONG, void *) = func; double *, BLASLONG, void *) = (void (*)(BLASLONG, BLASLONG, BLASLONG, double, double *, BLASLONG,
double *, BLASLONG, double *, BLASLONG, void *)) func;
afunc(args -> m, args -> n, args -> k, afunc(args -> m, args -> n, args -> k,
((double *)args -> alpha)[0], ((double *)args -> alpha)[0],
@ -220,7 +221,10 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
/* REAL / Single */ /* REAL / Single */
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, float, void (*afunc)(BLASLONG, BLASLONG, BLASLONG, float,
float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG,
float *, BLASLONG, void *) = func; float *, BLASLONG, void *) = (void (*)
(BLASLONG, BLASLONG, BLASLONG, float,
float *, BLASLONG, float *, BLASLONG,
float *, BLASLONG, void *)) func;
afunc(args -> m, args -> n, args -> k, afunc(args -> m, args -> n, args -> k,
((float *)args -> alpha)[0], ((float *)args -> alpha)[0],
@ -232,7 +236,9 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
/* REAL / BFLOAT16 */ /* REAL / BFLOAT16 */
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, bfloat16, void (*afunc)(BLASLONG, BLASLONG, BLASLONG, bfloat16,
bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG,
bfloat16 *, BLASLONG, void *) = func; bfloat16 *, BLASLONG, void *) = (void (*)(BLASLONG, BLASLONG, BLASLONG, bfloat16,
bfloat16 *, BLASLONG, bfloat16 *, BLASLONG,
bfloat16 *, BLASLONG, void *)) func;
afunc(args -> m, args -> n, args -> k, afunc(args -> m, args -> n, args -> k,
((bfloat16 *)args -> alpha)[0], ((bfloat16 *)args -> alpha)[0],
@ -243,7 +249,9 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
/* REAL / BLAS_STOBF16 */ /* REAL / BLAS_STOBF16 */
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, float, void (*afunc)(BLASLONG, BLASLONG, BLASLONG, float,
float *, BLASLONG, bfloat16 *, BLASLONG, float *, BLASLONG, bfloat16 *, BLASLONG,
float *, BLASLONG, void *) = func; float *, BLASLONG, void *) = (void (*)(BLASLONG, BLASLONG, BLASLONG, float,
float *, BLASLONG, bfloat16 *, BLASLONG,
float *, BLASLONG, void *)) func;
afunc(args -> m, args -> n, args -> k, afunc(args -> m, args -> n, args -> k,
((float *)args -> alpha)[0], ((float *)args -> alpha)[0],
@ -254,7 +262,9 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
/* REAL / BLAS_DTOBF16 */ /* REAL / BLAS_DTOBF16 */
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, double, void (*afunc)(BLASLONG, BLASLONG, BLASLONG, double,
double *, BLASLONG, bfloat16 *, BLASLONG, double *, BLASLONG, bfloat16 *, BLASLONG,
double *, BLASLONG, void *) = func; double *, BLASLONG, void *) = (void (*)(BLASLONG, BLASLONG, BLASLONG, double,
double *, BLASLONG, bfloat16 *, BLASLONG,
double *, BLASLONG, void *)) func;
afunc(args -> m, args -> n, args -> k, afunc(args -> m, args -> n, args -> k,
((double *)args -> alpha)[0], ((double *)args -> alpha)[0],
@ -271,7 +281,9 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
/* COMPLEX / Extended Double */ /* COMPLEX / Extended Double */
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble, void (*afunc)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble,
xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG,
xdouble *, BLASLONG, void *) = func; xdouble *, BLASLONG, void *) = (void (*)(BLASLONG, BLASLONG, BLASLONG, xdouble, xdouble,
xdouble *, BLASLONG, xdouble *, BLASLONG,
xdouble *, BLASLONG, void *)) func;
afunc(args -> m, args -> n, args -> k, afunc(args -> m, args -> n, args -> k,
((xdouble *)args -> alpha)[0], ((xdouble *)args -> alpha)[0],
@ -285,7 +297,9 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
/* COMPLEX / Double */ /* COMPLEX / Double */
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, double, double, void (*afunc)(BLASLONG, BLASLONG, BLASLONG, double, double,
double *, BLASLONG, double *, BLASLONG, double *, BLASLONG, double *, BLASLONG,
double *, BLASLONG, void *) = func; double *, BLASLONG, void *) = (void (*)(BLASLONG, BLASLONG, BLASLONG, double, double,
double *, BLASLONG, double *, BLASLONG,
double *, BLASLONG, void *)) func;
afunc(args -> m, args -> n, args -> k, afunc(args -> m, args -> n, args -> k,
((double *)args -> alpha)[0], ((double *)args -> alpha)[0],
@ -297,7 +311,9 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
/* COMPLEX / Single */ /* COMPLEX / Single */
void (*afunc)(BLASLONG, BLASLONG, BLASLONG, float, float, void (*afunc)(BLASLONG, BLASLONG, BLASLONG, float, float,
float *, BLASLONG, float *, BLASLONG, float *, BLASLONG, float *, BLASLONG,
float *, BLASLONG, void *) = func; float *, BLASLONG, void *) = (void (*)(BLASLONG, BLASLONG, BLASLONG, float, float,
float *, BLASLONG, float *, BLASLONG,
float *, BLASLONG, void *)) func;
afunc(args -> m, args -> n, args -> k, afunc(args -> m, args -> n, args -> k,
((float *)args -> alpha)[0], ((float *)args -> alpha)[0],
@ -425,7 +441,7 @@ blas_queue_t *tscq;
#endif #endif
if (queue) { if (queue) {
int (*routine)(blas_arg_t *, void *, void *, void *, void *, BLASLONG) = queue -> routine; int (*routine)(blas_arg_t *, void *, void *, void *, void *, BLASLONG) = (int (*)(blas_arg_t *, void *, void *, void *, void *, BLASLONG))queue -> routine;
atomic_store_queue(&thread_status[cpu].queue, (blas_queue_t *)1); atomic_store_queue(&thread_status[cpu].queue, (blas_queue_t *)1);
@ -503,7 +519,7 @@ blas_queue_t *tscq;
legacy_exec(routine, queue -> mode, queue -> args, sb); legacy_exec(routine, queue -> mode, queue -> args, sb);
} else } else
if (queue -> mode & BLAS_PTHREAD) { if (queue -> mode & BLAS_PTHREAD) {
void (*pthreadcompat)(void *) = queue -> routine; void (*pthreadcompat)(void *) = (void(*)(void*))queue -> routine;
(pthreadcompat)(queue -> args); (pthreadcompat)(queue -> args);
} else } else
(routine)(queue -> args, queue -> range_m, queue -> range_n, sa, sb, queue -> position); (routine)(queue -> args, queue -> range_m, queue -> range_n, sa, sb, queue -> position);
@ -871,13 +887,13 @@ int exec_blas(BLASLONG num, blas_queue_t *queue){
fprintf(STDERR, "\n"); fprintf(STDERR, "\n");
#endif #endif
routine = queue -> routine; routine = (int (*)(blas_arg_t *, void *, void *, double *, double *, BLASLONG))queue -> routine;
if (queue -> mode & BLAS_LEGACY) { if (queue -> mode & BLAS_LEGACY) {
legacy_exec(routine, queue -> mode, queue -> args, queue -> sb); legacy_exec(routine, queue -> mode, queue -> args, queue -> sb);
} else } else
if (queue -> mode & BLAS_PTHREAD) { if (queue -> mode & BLAS_PTHREAD) {
void (*pthreadcompat)(void *) = queue -> routine; void (*pthreadcompat)(void *) = (void (*)(void*))queue -> routine;
(pthreadcompat)(queue -> args); (pthreadcompat)(queue -> args);
} else } else
(routine)(queue -> args, queue -> range_m, queue -> range_n, (routine)(queue -> args, queue -> range_m, queue -> range_n,

2
interface/axpy.c
View File

@ -115,7 +115,7 @@ void CNAME(blasint n, FLOAT alpha, FLOAT *x, blasint incx, FLOAT *y, blasint inc
#endif #endif
blas_level1_thread(mode, n, 0, 0, &alpha, blas_level1_thread(mode, n, 0, 0, &alpha,
x, incx, y, incy, NULL, 0, (void *)AXPYU_K, nthreads); x, incx, y, incy, NULL, 0, (int (*)(void))AXPYU_K, nthreads);
} }
#endif #endif

2
interface/scal.c
View File

@ -102,7 +102,7 @@ void CNAME(blasint n, FLOAT alpha, FLOAT *x, blasint incx){
#else #else
&alpha, &alpha,
#endif #endif
x, incx, NULL, 0, NULL, 0, (void *)SCAL_K, nthreads); x, incx, NULL, 0, NULL, 0, (int (*)(void))SCAL_K, nthreads);
} }
#endif #endif

4
interface/zaxpy.c
View File

@ -128,9 +128,9 @@ void CNAME(blasint n, FLOAT *ALPHA, FLOAT *x, blasint incx, FLOAT *y, blasint in
blas_level1_thread(mode, n, 0, 0, ALPHA, x, incx, y, incy, NULL, 0, blas_level1_thread(mode, n, 0, 0, ALPHA, x, incx, y, incy, NULL, 0,
#ifndef CONJ #ifndef CONJ
(void *)AXPYU_K, (int (*)(void))AXPYU_K,
#else #else
(void *)AXPYC_K, (int (*)(void))AXPYC_K,
#endif #endif
nthreads); nthreads);
} }

2
interface/zscal.c
View File

@ -108,7 +108,7 @@ void CNAME(blasint n, FLOAT alpha_r, void *vx, blasint incx){
mode = BLAS_SINGLE | BLAS_COMPLEX; mode = BLAS_SINGLE | BLAS_COMPLEX;
#endif #endif
blas_level1_thread(mode, n, 0, 0, alpha, x, incx, NULL, 0, NULL, 0, (void *)SCAL_K, nthreads); blas_level1_thread(mode, n, 0, 0, alpha, x, incx, NULL, 0, NULL, 0, (int (*)(void))SCAL_K, nthreads);
} }
#endif #endif

2
kernel/x86_64/casum.c
View File

@ -130,7 +130,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
mode = BLAS_DOUBLE | BLAS_COMPLEX; mode = BLAS_DOUBLE | BLAS_COMPLEX;
#endif #endif
blas_level1_thread_with_return_value(mode, n, 0, 0, dummy_alpha, x, inc_x, blas_level1_thread_with_return_value(mode, n, 0, 0, dummy_alpha, x, inc_x,
NULL, 0, result, 0, (void *)asum_thread_function, nthreads); NULL, 0, result, 0, (int (*)(void))asum_thread_function, nthreads);
ptr = (FLOAT *)result; ptr = (FLOAT *)result;
for (i = 0; i < nthreads; i++) { for (i = 0; i < nthreads; i++) {
sumf += (*ptr); sumf += (*ptr);

2
kernel/x86_64/dasum.c
View File

@ -114,7 +114,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
#else #else
mode = BLAS_DOUBLE | BLAS_REAL; mode = BLAS_DOUBLE | BLAS_REAL;
#endif #endif
blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha, x, inc_x, NULL, 0, result, 0, (void *)asum_thread_function, nthreads); blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha, x, inc_x, NULL, 0, result, 0, (int (*)(void))asum_thread_function, nthreads);
ptr = (FLOAT *)result; ptr = (FLOAT *)result;
for (i = 0; i < nthreads; i++) { for (i = 0; i < nthreads; i++) {
sumf += (*ptr); sumf += (*ptr);

2
kernel/x86_64/ddot.c
View File

@ -190,7 +190,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
#endif #endif
blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha, blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha,
x, inc_x, y, inc_y, result, 0, x, inc_x, y, inc_y, result, 0,
( void *)dot_thread_function, nthreads); (int (*)(void)) dot_thread_function, nthreads);
ptr = (RETURN_TYPE *)result; ptr = (RETURN_TYPE *)result;
for (i = 0; i < nthreads; i++) { for (i = 0; i < nthreads; i++) {

2
kernel/x86_64/drot.c
View File

@ -196,7 +196,7 @@ int CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT
#else #else
int mode = BLAS_SINGLE | BLAS_REAL | BLAS_PTHREAD; int mode = BLAS_SINGLE | BLAS_REAL | BLAS_PTHREAD;
#endif #endif
blas_level1_thread(mode, n, 0, 0, alpha, x, inc_x, y, inc_y, &dummy_c, 0, (void *)rot_thread_function, nthreads); blas_level1_thread(mode, n, 0, 0, alpha, x, inc_x, y, inc_y, &dummy_c, 0, (int (*)(void))rot_thread_function, nthreads);
} }
#else #else
rot_compute(n, x, inc_x, y, inc_y, c, s); rot_compute(n, x, inc_x, y, inc_y, c, s);

2
kernel/x86_64/sasum.c
View File

@ -123,7 +123,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
#else #else
mode = BLAS_DOUBLE | BLAS_REAL; mode = BLAS_DOUBLE | BLAS_REAL;
#endif #endif
blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha, x, inc_x, NULL, 0, result, 0, (void *)asum_thread_function, nthreads); blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha, x, inc_x, NULL, 0, result, 0, (int (*)(void))asum_thread_function, nthreads);
ptr = (FLOAT *)result; ptr = (FLOAT *)result;
for (i = 0; i < nthreads; i++) { for (i = 0; i < nthreads; i++) {
sumf += (*ptr); sumf += (*ptr);

2
kernel/x86_64/srot.c
View File

@ -198,7 +198,7 @@ int CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT
#else #else
int mode = BLAS_SINGLE | BLAS_REAL | BLAS_PTHREAD; int mode = BLAS_SINGLE | BLAS_REAL | BLAS_PTHREAD;
#endif #endif
blas_level1_thread(mode, n, 0, 0, alpha, x, inc_x, y, inc_y, &dummy_c, 0, (void *)rot_thread_function, nthreads); blas_level1_thread(mode, n, 0, 0, alpha, x, inc_x, y, inc_y, &dummy_c, 0, (int (*)(void))rot_thread_function, nthreads);
} }
#else #else
rot_compute(n, x, inc_x, y, inc_y, c, s); rot_compute(n, x, inc_x, y, inc_y, c, s);

2
kernel/x86_64/zasum.c
View File

@ -130,7 +130,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
mode = BLAS_DOUBLE | BLAS_COMPLEX; mode = BLAS_DOUBLE | BLAS_COMPLEX;
#endif #endif
blas_level1_thread_with_return_value(mode, n, 0, 0, dummy_alpha, x, inc_x, blas_level1_thread_with_return_value(mode, n, 0, 0, dummy_alpha, x, inc_x,
NULL, 0, result, 0, (void *)asum_thread_function, nthreads); NULL, 0, result, 0, (int (*)(void))asum_thread_function, nthreads);
ptr = (FLOAT *)result; ptr = (FLOAT *)result;
for (i = 0; i < nthreads; i++) { for (i = 0; i < nthreads; i++) {
sumf += (*ptr); sumf += (*ptr);

2
kernel/x86_64/zdot.c
View File

@ -215,7 +215,7 @@ OPENBLAS_COMPLEX_FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLA
blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha, blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha,
x, inc_x, y, inc_y, result, 0, x, inc_x, y, inc_y, result, 0,
( void *)zdot_thread_function, nthreads); (int (*)(void))zdot_thread_function, nthreads);
ptr = (OPENBLAS_COMPLEX_FLOAT *)result; ptr = (OPENBLAS_COMPLEX_FLOAT *)result;
for (i = 0; i < nthreads; i++) { for (i = 0; i < nthreads; i++) {

2
lapack/getrf/getrf_parallel.c
View File

@ -662,7 +662,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
blas_level1_thread(mode, bk, is + bk + offset + 1, mn + offset, (void *)dummyalpha, blas_level1_thread(mode, bk, is + bk + offset + 1, mn + offset, (void *)dummyalpha,
a + (- offset + is * lda) * COMPSIZE, lda, NULL, 0, a + (- offset + is * lda) * COMPSIZE, lda, NULL, 0,
ipiv, 1, (void *)LASWP_PLUS, args -> nthreads); ipiv, 1, (int (*)(void))LASWP_PLUS, args -> nthreads);
is += bk; is += bk;
} }

4
lapack/lauum/lauum_L_parallel.c
View File

@ -102,7 +102,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
newarg.c = a; newarg.c = a;
syrk_thread(mode | BLAS_TRANSA_T | BLAS_TRANSB_N | BLAS_UPLO, syrk_thread(mode | BLAS_TRANSA_T | BLAS_TRANSB_N | BLAS_UPLO,
&newarg, NULL, NULL, (void *)HERK_LC, sa, sb, args -> nthreads); &newarg, NULL, NULL, (int (*)(void))HERK_LC, sa, sb, args -> nthreads);
newarg.m = bk; newarg.m = bk;
newarg.n = i; newarg.n = i;
@ -110,7 +110,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
newarg.b = a + (i ) * COMPSIZE; newarg.b = a + (i ) * COMPSIZE;
gemm_thread_n(mode | BLAS_TRANSA_T, gemm_thread_n(mode | BLAS_TRANSA_T,
&newarg, NULL, NULL, (void *)TRMM_LCLN, sa, sb, args -> nthreads); &newarg, NULL, NULL, (int (*)(void))TRMM_LCLN, sa, sb, args -> nthreads);
newarg.m = bk; newarg.m = bk;
newarg.n = bk; newarg.n = bk;

4
lapack/lauum/lauum_U_parallel.c
View File

@ -102,7 +102,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
newarg.c = a; newarg.c = a;
syrk_thread(mode | BLAS_TRANSA_N | BLAS_TRANSB_T, syrk_thread(mode | BLAS_TRANSA_N | BLAS_TRANSB_T,
&newarg, NULL, NULL, (void *)HERK_UN, sa, sb, args -> nthreads); &newarg, NULL, NULL, (int (*)(void))HERK_UN, sa, sb, args -> nthreads);
newarg.m = i; newarg.m = i;
newarg.n = bk; newarg.n = bk;
@ -110,7 +110,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
newarg.b = a + ( i * lda) * COMPSIZE; newarg.b = a + ( i * lda) * COMPSIZE;
gemm_thread_m(mode | BLAS_TRANSA_T | BLAS_RSIDE, gemm_thread_m(mode | BLAS_TRANSA_T | BLAS_RSIDE,
&newarg, NULL, NULL, (void *)TRMM_RCUN, sa, sb, args -> nthreads); &newarg, NULL, NULL, (int (*)(void))TRMM_RCUN, sa, sb, args -> nthreads);
newarg.m = bk; newarg.m = bk;
newarg.n = bk; newarg.n = bk;

4
lapack/potrf/potrf_L_parallel.c
View File

@ -110,7 +110,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
newarg.b = a + (i + bk + i * lda) * COMPSIZE; newarg.b = a + (i + bk + i * lda) * COMPSIZE;
gemm_thread_m(mode | BLAS_RSIDE | BLAS_TRANSA_T | BLAS_UPLO, gemm_thread_m(mode | BLAS_RSIDE | BLAS_TRANSA_T | BLAS_UPLO,
&newarg, NULL, NULL, (void *)TRSM_RCLN, sa, sb, args -> nthreads); &newarg, NULL, NULL, (int (*)(void))TRSM_RCLN, sa, sb, args -> nthreads);
newarg.n = n - i - bk; newarg.n = n - i - bk;
newarg.k = bk; newarg.k = bk;
@ -121,7 +121,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
HERK_THREAD_LN(&newarg, NULL, NULL, sa, sb, 0); HERK_THREAD_LN(&newarg, NULL, NULL, sa, sb, 0);
#else #else
syrk_thread(mode | BLAS_TRANSA_N | BLAS_TRANSB_T | BLAS_UPLO, syrk_thread(mode | BLAS_TRANSA_N | BLAS_TRANSB_T | BLAS_UPLO,
&newarg, NULL, NULL, (void *)HERK_LN, sa, sb, args -> nthreads); &newarg, NULL, NULL, (int (*)(void))HERK_LN, sa, sb, args -> nthreads);
#endif #endif
} }
} }

4
lapack/potrf/potrf_U_parallel.c
View File

@ -110,7 +110,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
newarg.b = a + (i + (i + bk) * lda) * COMPSIZE; newarg.b = a + (i + (i + bk) * lda) * COMPSIZE;
gemm_thread_n(mode | BLAS_TRANSA_T, gemm_thread_n(mode | BLAS_TRANSA_T,
&newarg, NULL, NULL, (void *)TRSM_LCUN, sa, sb, args -> nthreads); &newarg, NULL, NULL, (int (*)(void))TRSM_LCUN, sa, sb, args -> nthreads);
newarg.n = n - i - bk; newarg.n = n - i - bk;
newarg.k = bk; newarg.k = bk;
@ -121,7 +121,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
HERK_THREAD_UC(&newarg, NULL, NULL, sa, sb, 0); HERK_THREAD_UC(&newarg, NULL, NULL, sa, sb, 0);
#else #else
syrk_thread(mode | BLAS_TRANSA_N | BLAS_TRANSB_T, syrk_thread(mode | BLAS_TRANSA_N | BLAS_TRANSB_T,
&newarg, NULL, NULL, (void *)HERK_UC, sa, sb, args -> nthreads); &newarg, NULL, NULL, (int (*)(void))HERK_UC, sa, sb, args -> nthreads);
#endif #endif
} }
} }